[AArch64] Preserve X8 for thunks ending in variadic musttail calls

[rnk]

Summary

On Windows, X8 may be used to pass in the address of an aggregate that
is returned indirectly. Therefore, it should be forwarded to variadic
musttail calls and preserved in thunks.

Fixes PR41997

[efriedma]

The varargs argument list for a function normally can't contain an sret parameter; I guess musttail vararg lists are special. Maybe worth calling out in the LangRef description of mustttail? (I guess you could argue that sret is an ABI issue which is inherently target-specific, but it still seems a little surprising.)

Otherwise LGTM

[rnk]

In D62344#1514876, @efriedma wrote:

The varargs argument list for a function normally can't contain an sret parameter; I guess musttail vararg lists are special. Maybe worth calling out in the LangRef description of mustttail? (I guess you could argue that sret is an ABI issue which is inherently target-specific, but it still seems a little surprising.)

There was half a sentence in langref about forwarding varargs, but it was really meant to be a stronger guarantee. I added another bullet there, feel free to take a look and suggest improvements.

Otherwise LGTM

Thanks.

[rnk]

In D62344#1515074, @rnk wrote:

There was half a sentence in langref about forwarding varargs, but it was really meant to be a stronger guarantee. I added another bullet there, feel free to take a look and suggest improvements.

This didn't make it into this commit, so I committed it in rL361590.

llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp

if (MFI.hasMustTailInVarArgFunc()) {
       SmallVector<MVT, 2> RegParmTypes;
       RegParmTypes.push_back(MVT::i64);
       RegParmTypes.push_back(MVT::f128);
       // Compute the set of forwarded registers. The rest are scratch.
       SmallVectorImpl<ForwardedRegister> &Forwards =
                                        FuncInfo->getForwardedMustTailRegParms();
       CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes,
                                                CC_AArch64_AAPCS);
+
+      // Conservatively forward X8, since it might be used for aggregate return.
+      if (!CCInfo.isAllocated(AArch64::X8)) {
+        unsigned X8VReg = MF.addLiveIn(AArch64::X8, &AArch64::GPR64RegClass);
+        Forwards.push_back(ForwardedRegister(X8VReg, AArch64::X8, MVT::i64));
+      }
     }
   }
 
   // On Windows, InReg pointers must be returned, so record the pointer in a
   // virtual register at the start of the function so it can be returned in the
   // epilogue.
   if (IsWin64) {
     for (unsigned I = 0, E = Ins.size(); I != E; ++I) {

llvm/trunk/test/CodeGen/AArch64/vararg-tallcall.ll

entry:
   ret void
                                                   ; No predecessors!
   ret void
 }
 
 attributes #1 = { noinline optnone "thunk" }
 
 ; CHECK: mov     v16.16b, v0.16b
-; CHECK: ldr     x8, [x0]
-; CHECK: ldr     x8, [x8]
+; CHECK: ldr     x9, [x0]
+; CHECK: ldr     x9, [x9]
 ; CHECK: mov     v0.16b, v16.16b
-; CHECK: br      x8
+; CHECK: br      x9